OpenAI Releases Two Powerful Open-Source Reasoning Models

In a significant shift from its typically closed-source strategy, OpenAI on Tuesday launched two open-weight AI reasoning models, aiming to challenge the top performers in the open-source AI space. The models—gpt-oss-120b and gpt-oss-20b—are now freely available for download on Hugging Face, OpenAI confirmed, and come with a permissive Apache 2.0 license, allowing both commercial use and modification.

The move marks OpenAI’s first open release of a language model since GPT-2 in 2019, signaling a possible pivot in its openness philosophy and perhaps a response to increasing competition from global AI developers—particularly those based in China.

Model Specs: Big Power, Broad Accessibility

The two models come in different sizes and capabilities:

GPT-OSS-120B is the larger and more powerful of the two, boasting 117 billion parameters. Despite its scale, it can run on a single high-end Nvidia GPU, thanks to a Mixture-of-Experts (MoE) architecture that activates just 5.1 billion parameters per token.
GPT-OSS-20B is a lighter model designed for accessibility—it can run on consumer-grade hardware, such as a laptop with 16 GB of RAM.

OpenAI's Reasoning model — Image Credit: OpenAI

This tiered approach enables a wide range of developers to work with cutting-edge models, from hobbyists to enterprise teams.

A Return to “Open” Roots—With Limits

OpenAI was once known for championing open research and open-source tools, but that ethos faded as the company began scaling its powerful GPT models behind an API paywall. However, CEO Sam Altman has publicly reflected on this change of course. Earlier this year, he acknowledged that OpenAI had been “on the wrong side of history” regarding open-sourcing.

Now, with GPT-oss, the company appears to be trying to strike a balance—sharing high-performing models with the public while keeping training datasets closed, citing safety and legal concerns.

“To that end, we are excited for the world to be building on an open AI stack created in the United States, based on democratic values, available for free to all and for wide benefit,” Altman said in a statement.

The move also aligns with recent political pressure from the Trump administration, which has encouraged American AI companies to open source more of their technologies to counter China’s influence in global AI development.

Competitive Edge: How Well Do These Models Perform?

OpenAI claims its open models outperform most of the competition across several benchmark tests, even though they still fall short of its proprietary o-series (like o3 and o4-mini).

On Codeforces (with tools) — a competitive programming benchmark:

gpt-oss-120b scored 2622
gpt-oss-20b scored 2516
Both outperformed DeepSeek’s R1, though still trailed OpenAI’s internal models.

On Humanity’s Last Exam (HLE) — a diverse question set measuring general reasoning:

gpt-oss-120b scored 19%
gpt-oss-20b scored 17.3%
Again, both models surpassed leading open alternatives like DeepSeek and Qwen, while falling behind OpenAI’s o3.

But there are trade-offs. OpenAI’s gpt-oss models showed significantly higher hallucination rates than its closed models. In the PersonQA benchmark:

gpt-oss-120b hallucinated on 49% of queries
gpt-oss-20b hallucinated on 53%

By contrast, OpenAI’s o1 model hallucinated just 16% of the time, and o4-mini came in at 36%. The company attributed this difference to the smaller scale and limited “world knowledge” of open models, which tend to fabricate more when asked factual or complex questions.

Training and Architecture: Efficiency and Tool Use

The open models were trained using similar techniques to OpenAI’s o-series, including reinforcement learning (RL) and chain-of-thought reasoning. They’re also designed to power AI agents by interacting with external tools like Python, web search, or APIs during their reasoning process.

While text-only in their current form (no image or audio processing), the models can delegate complex queries to cloud-based AI models—creating hybrid workflows where open models serve as orchestrators for more advanced, closed systems.

Legal and Ethical Boundaries: No Training Data Released

While the Apache 2.0 license makes these models free for commercial use, OpenAI is not releasing the training data behind gpt-oss. This is a clear legal safeguard. Multiple lawsuits against OpenAI and other AI companies are challenging the legality of using copyrighted material to train models, and transparency around datasets has become a legal risk.

Even with those precautions, OpenAI delayed the release multiple times to address safety evaluations, including concerns about whether bad actors could fine-tune the models for misuse, such as in cyberattacks or biological threat simulations. The company found no evidence that GPT-oss could reach high-danger thresholds in these areas, though it did note marginal increases in biological knowledge.

The Geopolitical Angle: Competing with China’s Rise

The timing of this release is no coincidence. Chinese labs like DeepSeek, Qwen (Alibaba), and Moonshot AI have made rapid advances in the open model space. While Meta once dominated the open frontier with LLaMA, it has recently retreated from full openness, citing concerns about misuse—leaving room for challengers.

By launching a state-of-the-art open model under U.S. leadership, OpenAI is not just targeting developers—it’s making a geopolitical statement.

What’s Next?

With gpt-oss-120b and gpt-oss-20b now in the wild, attention turns to what’s next. Developers are eagerly awaiting the debut of DeepSeek R2 and Meta’s rumored open model from its Superintelligence Lab.

For now, OpenAI has reclaimed some of the open-source spotlight. The question is whether it can maintain developer trust and momentum without releasing its training data—and whether these open models are truly a step toward more democratized AI or simply a clever strategic move in the cloud and model wars.